QryptaShield: Quantum-Inspired Feature Selection and RL-Calibrated Federated Learning for Low-False-Positive IoT Intrusion Detection

Abstract

The exponential growth of heterogeneous IoT edge data has intensified challenges in intrusion detection due to high dimensionality, class imbalance, and privacy constraints, while fixed thresholds and centralized training increase false positives and risk data leakage. To address these issues, this study proposes QryptaShield, an end-to-end intrusion detection framework unifying Quantum-Informed Adaptive Feature Selection (QIAFS), Neuro-Swarm Edge AI (NSEA), quantum-Augmented Reinforcement learning (QARL), and a Zero-Knowledge Secure Communication Layer (ZKSCL). QIAFS leverages entanglement-aware ranking and quantum kernels to reduce redundancy, NSEA enables lightweight on-device CNN inference with swarm-based adaptation, QARL dynamically adjusts thresholds and defense actions via quantum-enhanced RL, and ZKSCL ensures privacy-preserving, trust-weighted federated aggregation through lattice-based homomorphic encryption and blockchain-backed trust scoring. Evaluated against classical baselines including Logistic Regression, SVM, Naïve Bayes, kNN, AdaBoost, and RFF-based models, the proposed ensemble achieves Accuracy = 92.4%, F1 = 0.902, AUC = 0.976, and AP = 0.933, with FPR = 0.053 and Recall = 0.889, outperforming or matching the best classical baselines while significantly lowering false positives. Results confirm that QryptaShield delivers scalable, interpretable, and privacy-preserving IoT intrusion detection, offering a practical pathway toward next-generation anomaly detection and fraud analytics in resource-constrained, adversarial environments.